SALSA: Structure-Aware LoRA for Summarizing Agendas

Meeting summarization is a critical task in organizational knowledge management, yet existing approaches suffer from two fundamental limitations: naive token-level truncation that breaks semantic coherence, and computationally expensive full finetuning impractical for resource-constrained deployments. This paper presents SALSA (Structure-Aware LoRA for Summarizing Agendas), a dual-modification framework addressing both limitations simultaneously on the MeetingBank benchmark. SALSA combines (1) sentence-boundary-aware chunking with sliding overlap, and (2) selective LoRA on cross-attention layers, reducing trainable parameters to 0.21% of the full model (294K vs. 139.7M). A 2x2 ablation study demonstrates two key findings: sentence-aware chunking maintains performance equivalent to full fine-tuning (ROUGE-1: 65.92 vs. 65.48), and chunking substantially improves LoRA effectiveness by +6.69 ROUGE-1 points over LoRA-only baseline. These findings establish SALSA as a practical approach for efficient organizational meeting summarization